Tesla Optimus Gen 3 Architecture: Compute, Battery, Actuators

The Tesla Optimus Gen 3 architecture is the first humanoid robot platform to ship a vertically integrated stack where the inference SoC, actuators, battery cells, and Vision-Language-Action (VLA) policy are all designed by the same company. That is the original thesis of this post: Optimus Gen 3 is not the most capable humanoid in any single dimension, but it is the only one whose every layer is co-designed against a fleet that already has thousands of hours of work in real Tesla factories. This reference architecture deep-dive walks the full Tesla Optimus Gen 3 architecture top to bottom — onboard compute, the in-house electromechanical actuators, the ~2.3 kWh torso battery, the vision-only sensing stack, and the VLA software trained against Tesla’s Dojo cluster — and then compares it head-to-head with Figure 02, 1X NEO, and the electric Atlas. Where Tesla has only hinted at a number, I label it as an estimate from teardowns and Tesla AI Day disclosures. Where the number is published, I cite it.

Architecture at a glance

Why Optimus Gen 3 matters in 2026

Optimus Gen 3 matters because it is the first general-purpose humanoid to be deployed inside the customer’s own factories at fleet scale before any external sale. That gives Tesla a closed perceive-act-label-retrain loop that competitors cannot match, and it forces the architecture to optimize for cost-per-hour-of-useful-work, not for demo videos.

Humanoids have been a research project for thirty years. What changed in 2024-2026 is three things at once. First, transformer-based VLA models such as Physical Intelligence’s pi-0 and Google’s RT-2 made it tractable to map raw camera frames to joint commands without hand-coded motion primitives. Second, lithium-ion energy density crossed the threshold (>260 Wh/kg at cell level for high-nickel NMC) where a 2 kWh pack could fit in a torso and still deliver a full work shift. Third, the cost of harmonic-drive actuators collapsed once Tesla, Figure, and 1X all started designing in-house. Tesla disclosed at the Tesla AI Day 2022 presentation that Optimus would use bespoke actuators with explicit force, position, and current sensing — Gen 3 is the third iteration of that program. The Gen 3 spec sheet still has gaps, but the architecture has stabilized enough to talk about as a reference design.

Tesla Optimus Gen 3 architecture: the full system view

The Tesla Optimus Gen 3 architecture is a five-layer stack: sensing, compute, software, actuation, and power, with a thin connectivity layer hanging off the side. Sensors push into a single SoC that hosts perception, the VLA policy, and the motion planner. The planner emits torque setpoints that an FPGA-class inner loop turns into PWM commands at 1 kHz, gated by an ASIL-class safety MCU.

The reference architecture is intentionally similar to Tesla’s vehicle stack. The perception network is a direct descendant of Tesla’s Hydranet multi-task vision backbone used in FSD — same image rectification approach, same occupancy-network style 3D output, retrained on humanoid-mounted camera streams. The safety MCU pattern (one performance compute domain plus a separate certifiable domain that can disable the actuators) is lifted almost unchanged from the vehicle. Even the BMS firmware shares lineage with Tesla’s automotive packs, just scaled to ~2.3 kWh.

That vertical reuse is the single biggest architectural choice in the platform. Figure has to negotiate with NVIDIA for Jetson Thor allocation. 1X integrates third-party harmonic drives. Boston Dynamics under Hyundai relies on Hyundai Mobis for power electronics. Tesla owns silicon, motors, cells, and the data center training the policy. The cost is generality — none of these subsystems are designed for any robot other than Optimus — and the payoff is a tight thermal, mechanical, and software envelope that an integrator could not assemble from off-the-shelf parts.

Below I walk each layer with the published numbers, the estimated numbers, and the architectural decision behind each.

Onboard compute: a Tesla AI inference SoC derivative

Optimus Gen 3 runs on a single Tesla-designed inference SoC in the ~500 TOPS class, paired with an ASIL-D class safety MCU and a smaller power MCU that owns the battery management system. The performance domain hosts perception, VLA inference, and the whole-body MPC; the safety domain has authority to drop the actuator bridges within one control cycle if a hard fault is detected.

Tesla has not published the exact die for Optimus Gen 3. What is known publicly: at AI Day 2022 Tesla showed a board with what was clearly a derivative of the FSD HW3/HW4 inference architecture, the same dual-NPU design that runs the FSD stack in vehicles. The credible estimate, based on the published FSD HW4 figures and teardown chatter reported by The Information’s robotics coverage, puts the Optimus Gen 3 SoC in the 400-600 TOPS INT8 range with on-die SRAM in the tens of MB and shared LPDDR5X off-chip. Treat any tighter number as speculation.

The architectural decision worth explaining is the single-SoC choice. Most humanoid teams in 2025-2026 carry a GPU board plus a separate real-time MCU. The Jetson Thor reference, which I covered in detail in the NVIDIA Jetson Thor humanoid robot reference architecture, takes this dual-die path. Tesla collapses it. One SoC runs everything except the ASIL safety functions and the BMS. The justification is power and thermal: a humanoid torso has roughly 30-40 W of sustainable cooling budget for compute (the rest of the ~150 W thermal envelope is for actuators and power electronics), and two boards burn that twice over for PHYs, links, and idle power.

The trade-off is brutal. If the SoC faults, the robot has no fallback compute. Tesla’s answer is a watchdog architecture: the safety MCU monitors a 1 kHz heartbeat from the controller process running on the SoC; if two consecutive heartbeats are missed (2 ms), the safety MCU drives all H-bridges to high-impedance, the brakes engage, and the robot enters a passive-hold posture. There is no graceful degradation of autonomy — it is full autonomy or safe-stop.

/* Safety MCU watchdog pseudocode — illustrative, not Tesla source */
#define WD_PERIOD_US     1000
#define WD_MISS_LIMIT    2

static volatile uint32_t last_heartbeat_us;
static volatile uint16_t missed = 0;

void on_heartbeat_irq(uint32_t now_us) {
    last_heartbeat_us = now_us;
    missed = 0;
}

void on_timer_1khz(uint32_t now_us) {
    if ((now_us - last_heartbeat_us) > WD_PERIOD_US) {
        if (++missed >= WD_MISS_LIMIT) {
            actuator_bridges_disable();   /* gate drivers off */
            brakes_engage();              /* fail-safe brake */
            log_event(EVT_WATCHDOG_TRIP);
        }
    }
}

That is the contract. The SoC has to keep talking every millisecond or it loses its arms and legs. It is the same pattern Tesla uses in FSD between the main compute and the vehicle controller.

Thermal envelope and clock gating

Public teardown photos of Gen 2 showed the compute board sitting directly above the torso battery with a vapor-chamber heat spreader to a chest-mounted radiator. Gen 3 is believed to use the same approach. The architectural consequence is that VLA inference cannot run at peak TOPS continuously — the SoC throttles based on chassis temperature and battery temperature simultaneously. Tesla’s policy code is therefore designed around an action-chunking strategy (predict 16 actions at once, execute open-loop for ~320 ms at 50 Hz) so the SoC can burst-inference and idle, rather than holding 100% utilization. This is the same trick used in the Physical Intelligence pi-0 and Google RT-2 family of VLA foundation models, and Optimus appears to have converged on the same answer for thermal reasons.

Actuator design: in-house electromechanical, not hydraulic

Optimus Gen 3 uses 28 in-house electromechanical actuators split between rotary harmonic-drive units for high-torque joints (hip, shoulder) and roller-screw linear actuators for the limbs that need explosive force (knee extension, elbow). The two hands carry 22 additional tendon-driven degrees of freedom. There is no hydraulic actuation anywhere on the robot.

Tesla disclosed at AI Day 2022 six actuator families across the body. The Gen 3 design appears to consolidate to the same six families with updated torque envelopes. The published peak figures from Tesla’s own slides were ~250 Nm for the largest rotary unit and ~4 kN for the largest linear unit. Knee, ankle, and elbow take linear actuators because the lever arm geometry makes a roller-screw cheaper per Newton-meter of joint torque than a harmonic drive at the same mass.

Each actuator integrates four things in one housing: a permanent-magnet synchronous motor, a position encoder (absolute, hall-effect or magnetic), a strain-gauge-based torque sensor, and a current-sensing power stage. That four-signal-per-joint topology is what makes whole-body torque control possible at 1 kHz. Most legacy industrial arms have position-only feedback at the joint and infer torque from motor current, which is fine for repetitive pick-and-place but useless for contact-rich manipulation. Optimus measures torque directly at the output side of the reducer.

Tesla has never used the word “cycloidal” publicly for Optimus, despite enthusiast speculation. The visible Gen 2 hip joint photos look like strain-wave (harmonic-drive class) reducers with Tesla’s own gear cutting. Treat any Tesla cycloidal claim as unconfirmed.

Why electromechanical wins over hydraulic in 2026

Boston Dynamics’ original Atlas used hydraulics and could throw 50 kg of weight around with theatrical violence. The electric Atlas, unveiled in 2024, dropped hydraulics entirely. Tesla never considered them. The reasons are well documented in the IEEE Spectrum humanoid actuation coverage: hydraulic pumps are loud, leak, are hard to certify near humans, and have terrible energy efficiency end-to-end (a hydraulic Atlas burned roughly five times the battery of an equivalent electric design for the same task). Electromechanical actuators with high-ratio reducers and torque sensing get you 80% of the dynamic capability at 20% of the power.

The remaining 20% is real. An electromechanical knee cannot match a hydraulic knee for a single explosive jump. Optimus is not designed to do parkour. It is designed to walk between machines in a factory and lift parts.

Hand and tactile design

The Gen 3 hands are the most visibly upgraded subsystem versus Gen 2. They are tendon-driven (cables routed through the forearm pulled by linear actuators), with 22 DoF across both hands, and they carry tactile sensor arrays on the fingertips and palms. Tesla has not published the tactile array resolution, but visible Gen 3 demo footage shows the hands successfully passing in-hand re-grasping tasks (rotating an egg, threading a screw) that require dense contact feedback. A reasonable estimate based on competitor disclosures is 6-12 taxels per fingertip and ~30 across the palm, with capacitive or piezoresistive sensing at 100-200 Hz.

The choice of tendon drive over geared finger joints is a weight argument. Pulling tendons from forearm-mounted actuators keeps the hand mass under 500 g per side, which is what makes the wrist actuator small enough not to dominate forearm volume. The cost is backlash and tendon stretch — both are calibrated out in software with a per-tendon load cell, but it is the single most fiddly part of the robot to manufacture.

Battery: ~2.3 kWh NMC, torso-integrated

The Optimus Gen 3 battery is a ~2.3 kWh nickel-manganese-cobalt (NMC) pack integrated into the torso, with onboard battery management that is a direct lineage of Tesla’s automotive BMS. Published Tesla disclosures state a runtime target of approximately five hours of “useful work” or up to 22 hours of idle/standby on a single charge.

The 2.3 kWh figure was disclosed by Tesla at AI Day 2022 for the Gen 1 spec and has not been revised publicly for Gen 3. Cell chemistry is widely reported as cylindrical 21700 NMC; the cell count is undisclosed but ~50 cells in a 6S configuration is consistent with the form factor and a nominal pack voltage in the 50-54 V range. That voltage is what feeds the actuator inverters; a separate DC-DC steps it down to the compute rail.

Fast charging is the architectural choice that matters most for factory operations. Tesla has not published the Optimus charge rate, but the engineering target implied by the “5 hours work / 1 hour charge” duty cycle competitors are quoting is roughly 1.5-2 C, which is well within what high-nickel NMC will tolerate without significant cycle-life penalty. That implies an onboard charger or a docked charger in the 3-4 kW class. The pack is liquid-cooled because anything above 1 C without active cooling will exceed the 45 C cell-surface limit.

The thermal architecture is also lifted from automotive practice. Coolant loops the battery, the compute SoC, and the high-current power electronics, with a single chest-mounted radiator dumping heat to ambient. In a factory environment with ambient at 35 C this is the limiting factor — sustained heavy-lift tasks will trip thermal throttling on the SoC and reduce continuous torque on the largest actuators before the battery itself runs dry. Knowing this is what makes work-cell layout decisions different for humanoids than for fixed arms.

Trade-off: bigger battery vs. lighter robot

Every kilowatt-hour of battery adds roughly 4-5 kg of mass (cells + casing + BMS + cooling). On a humanoid with a target total mass under 60 kg, doubling the battery to 4.6 kWh would consume ~20 kg of payload and structural budget. Tesla chose the 2.3 kWh point and a fast-charge architecture instead. The implicit assumption is that Optimus will operate near a charging dock for most of its life, which is fine in a Tesla factory but a real constraint for any future field-service deployment.

Sensing: vision-first, LIDAR-free

Optimus Gen 3 sensing is vision-first — eight RGB cameras around the head and torso plus a torso IMU, joint encoders, hand-mounted tactile arrays, and a beamforming microphone array. There is no LIDAR, no time-of-flight depth camera, and no stereo depth module. All 3D scene understanding is inferred by the perception network from the 2D camera streams, the same philosophy Tesla applies to FSD.

This is a religious decision inside Tesla, not a cost decision. Elon Musk has stated repeatedly that the company believes humans drive and manipulate with vision alone, so vision is sufficient for robots if the neural networks are good enough. The architectural payoff is that the sensing bill of materials is dominated by cheap CMOS image sensors (likely Sony or OmniVision automotive-grade IMX-class parts) and a small set of ISPs. The cost is that occlusion handling and metric depth become the perception network’s problem, not the sensor’s.

The eight-camera layout is structured around three rings: a forward-looking pair on the head (for object manipulation and walking-path planning), four wider-angle units in the torso for 360-degree surround (for navigation and self-collision avoidance), and a downward-looking pair under the chin for foot-placement and hand-near-body tasks. Frame rates are not published, but to support 50 Hz VLA inference the cameras almost certainly run at 36 Hz or 60 Hz with hardware sync. Resolution per camera is estimated at 1280×960 based on the Hydranet lineage in FSD HW4.

Proprioception is the secret sauce

The most underweighted sensing modality on Optimus is proprioception — the joint-by-joint position, velocity, and torque measurements coming from every actuator at 4 kHz. Forty-plus axes at 4 kHz is roughly 160k samples per second of robot state, all timestamped and fed back to the controller. That is what allows zero-shot adaptation to slight model errors, payload mass changes, or floor compliance. Vision tells the robot where things are; proprioception tells the robot what it actually did.

This is the dimension where any home-grown humanoid built on third-party actuators (which usually expose only position) hits a wall. Tesla, Figure, and 1X all designed their actuators specifically to expose torque, and that is why their published manipulation demos look qualitatively different from research-lab robots assembled from off-the-shelf joints.

Software: VLA models trained on Tesla teleop + Dojo

The Optimus Gen 3 software stack is a Vision-Language-Action model running at 50 Hz on the onboard SoC, supplied with action chunks by a transformer-based policy that was trained on tens of thousands of hours of human teleoperation footage and refined on Tesla’s Dojo training cluster. Below the VLA, a whole-body model-predictive controller (MPC) solves a 50 ms-horizon quadratic program every 50 ms to convert end-effector trajectories into balanced joint torques, and a 1 kHz inner loop closes torque control on each actuator.

The data flow is the headline. Cameras capture at ~36 Hz, the ISP pushes rectified tensors into the perception network (a Hydranet-derived multi-task backbone producing occupancy, object bounding boxes, hand keypoints, semantic segmentation), perception tokens plus the current language goal go into the VLA transformer, the VLA emits a 16-step action chunk in roughly 30-40 ms, the chunk is unrolled into end-effector trajectories, the MPC turns those into joint torques, and the inner loop drives the actuators. End-to-end perceive-to-act latency budget is approximately 80 ms from photons to first torque change.

Teleop is the data engine

Tesla has been running teleoperation pods in Fremont and Austin for at least three years. Operators wear VR headsets and haptic gloves to puppet Optimus units doing real factory tasks. Every minute of teleop generates synchronized (camera frames, joint commands, success label) tuples that go straight into the training set. This is the single biggest reason Optimus’s manipulation quality outpaces academic research labs — Tesla has a data engine that competitors are still trying to bootstrap.

Choosing the right planner for the MPC layer is its own deep topic; if you are building a comparable stack on open-source tools, the MoveIt 2 vs Tesseract motion planning comparison walks through the trade-offs in detail. Tesla’s planner is custom and almost certainly closer to a Tesseract-style trajectory optimizer than to a sampling-based planner, but the published evidence is thin.

Comparison: Optimus vs Figure 02 vs 1X NEO vs Atlas

The comparison is summarized in the matrix above. Optimus wins on vertical integration and on data engine scale. Figure 02 wins on partnership velocity (BMW deployment is real and at scale). 1X NEO targets a different market entirely (home, soft tendon-driven, lower payload). The electric Atlas wins on raw dynamic envelope. None of them is clearly “the right” humanoid for any application yet; each is optimizing for a different bet about which segment becomes the first profitable use case.

Operating modes and fault handling

Optimus Gen 3 operates in five states: Boot/SelfCheck, Idle, Teleop, Autonomous, SafeStop, and a terminal Fault state. Transitions are mostly trivial — operator request, goal complete, dock detection — but the SafeStop entry conditions are worth understanding.

Any of the following triggers SafeStop within one control cycle: torque-limit breach on any actuator, contact-force fault detected by the F/T-sensing in the hand, low battery below 15% SoC, communications loss above 200 ms in teleop mode, or the safety MCU watchdog firing. In SafeStop the bridges drop, the brakes engage, posture is held passively by joint stiffness, and the last 5 seconds of state are flushed to the fleet API for triage. Recovery requires an explicit operator reset; the robot does not auto-resume.

The Fault state is terminal in software terms — it means the safety MCU has confirmed a hardware failure (sensor disagreement, BMS over-temperature, encoder fail) and the robot requires service. The fleet API receives a structured fault record; a human triage step decides repair vs. retire.

Trade-offs and failure modes

Vertical integration is the Optimus thesis, and it is also the Optimus failure mode. Several specific risks are visible from the architecture.

The single-SoC choice means there is no graceful degradation of autonomy. If the inference SoC throttles for thermal reasons under sustained heavy load, the VLA cycle stretches from 20 ms to 40 ms, the action chunk length grows, and the robot starts to look hesitant. There is no fallback to a “simpler” mode running on a secondary compute domain because there is no secondary compute domain. Tesla’s bet is that the SoC can hold its budget; the failure mode is that ambient heat in a real factory in July makes that bet fail.

Vision-only sensing has known failure modes that LIDAR-based stacks do not have. Specular reflections (polished metal panels on a chassis), low contrast (matte black parts on a black conveyor), and night-shift lighting changes have all been documented as failure cases for FSD and the same physics applies to Optimus. The mitigation is augmenting training data, not adding a sensor. That is a slow remediation loop.

Vendor lock to in-house actuators means there is no second source. If the Tesla actuator line at Lathrop has a quality escape, the whole Optimus production line stops. Figure can in principle dual-source. Tesla cannot.

The 2.3 kWh battery and 5-hour work runtime is sufficient for a one-shift duty cycle near a dock and insufficient for any field application longer than that. Outdoor inspection or remote logistics roles will need a different pack. Tesla has not signaled any intent to ship a larger-pack variant.

Teleop dependence is the under-discussed risk. The VLA quality is bounded by the quality of the teleop data set. Tesla’s data engine is excellent inside Tesla factories. The same model will not generalize zero-shot to a hospital, a restaurant, or a customer’s living room. Each new domain requires its own teleop campaign before the autonomy curve starts to bend.

Finally, the regulatory posture for humanoids near humans in commercial environments is unsettled. EN ISO 10218-1:2025 (industrial robots) does not cleanly apply, and the ISO TC 299 working group on personal-care humanoid safety is still in draft. Tesla is operating under what is essentially a research exemption inside its own factories. Commercial sale to third-party industrial customers will need a certifiable safety case that the Gen 3 architecture has not yet been audited for.

Manufacturing and deployment topology

Optimus Gen 3 ships into Tesla’s own factories first — Fremont and Austin — as a closed-loop deployment. Robots work cells alongside humans (pick-and-place, kitting, torque tasks), the on-robot edge MEC aggregates fleet telemetry, batched logs go to Tesla’s cloud data lake, the Dojo cluster plus an Isaac-style simulator re-train the VLA weekly, and signed firmware ships back to the fleet via an OTA channel with canary deployment.

The closed loop is the strategic moat. Every other humanoid vendor has to negotiate access to a customer’s factory to collect deployment data; Tesla owns the factory. Every robot-hour of work is a labeled training sample. The teleop pods supply recovery demonstrations any time the autonomous policy fails, those demos go back into the same lake, and the model improves on a weekly cadence. This is the same flywheel that made FSD viable at scale, applied to a different problem.

Practical recommendations

Specific guidance if you are evaluating Optimus or designing against a similar architecture:

• Do not benchmark Optimus by its dynamic envelope. It is not Atlas. Benchmark it by cost-per-hour-of-useful-work in a real cell with real cycle-time pressure.
• If you are building a competing humanoid, copy the four-signal-per-joint actuator pattern (position, current, output torque, temperature) before anything else. Without it, your VLA will plateau.
• Plan thermal budget around 30-40 W of sustainable compute and ~150 W total robot envelope. SoC throttling is the silent killer of demos.
• Treat the 2.3 kWh / 5-hour runtime as the design center, not the limit. Architect work cells around docking opportunities every 4-5 hours.
• Budget perceive-to-act latency at ~80 ms end-to-end and design tasks accordingly. Sub-50 ms reflex tasks should be hard-coded into the inner loop, not pushed through the VLA.
• For vision-only stacks, invest in dataset coverage of specular, low-contrast, and changing-illumination cases before you invest in more cameras.
• If you require a certifiable safety case (industrial customer, EU sale), assume 12-24 months of homologation work on top of the architecture itself.

FAQ

What compute chip does Tesla Optimus Gen 3 use?

Tesla has not formally published the Optimus Gen 3 inference chip name. Based on AI Day 2022 disclosures and credible teardown reporting from sources such as The Information, the SoC is a derivative of Tesla’s FSD HW3/HW4 inference architecture, in the ~500 TOPS INT8 class with on-die SRAM in the tens of MB. It hosts perception, the VLA policy, and the whole-body MPC together. A separate ASIL-class safety MCU and a power-management MCU handle real-time safety and BMS functions.

How long does Tesla Optimus Gen 3 run on a single battery charge?

Tesla disclosed an approximately 2.3 kWh torso battery for Optimus, targeting up to five hours of useful work or up to 22 hours of idle/standby per charge. Real runtime depends heavily on duty cycle: continuous heavy-lift tasks will exhaust the pack faster and may trigger thermal throttling first. The architecture is designed around a fast-charge dock cycle, so the practical metric is uptime per shift rather than peak runtime per charge.

Does Tesla Optimus use LIDAR?

No. Optimus Gen 3 uses a vision-only sensing stack — eight RGB cameras around the head and torso, plus joint encoders, an IMU, hand-mounted tactile arrays, and a microphone array. There is no LIDAR or dedicated depth camera. All 3D scene understanding is inferred by the perception network from 2D camera streams, the same philosophy Tesla applies to FSD. The trade-off is that occlusion and metric depth become a software problem rather than a sensor problem.

How does Tesla Optimus differ from Figure 02 and 1X NEO?

Tesla is vertically integrated end-to-end — SoC, actuators, cells, and the VLA training cluster are all in-house. Figure 02 uses third-party compute (NVIDIA + custom MCU) and targets enterprise partners such as BMW. 1X NEO uses tendon-driven compliant actuators, smaller battery, and targets the home market. Tesla wins on data engine scale because Optimus deploys into Tesla’s own factories first. Figure wins on partnership velocity. 1X wins on safety around humans.

What are Tesla Optimus actuators made of?

Each actuator integrates a permanent-magnet synchronous motor, a high-ratio reducer (harmonic-drive class for rotary joints, roller-screw for linear), a position encoder, a strain-gauge torque sensor, and an integrated current-sensing power stage. Twenty-eight large actuators serve the body across six design families; the hands are tendon-driven with linear actuators in the forearm pulling cables to 22 hand degrees of freedom. There is no hydraulic actuation anywhere on Gen 3.

When will Tesla Optimus be available for purchase?

Tesla has publicly stated an intent to sell Optimus to external customers, but as of 2026 the fleet is operating inside Tesla’s own factories. External sales require a certifiable safety case under emerging standards (ISO TC 299 personal-care robot work, EN ISO 10218 successor standards for industrial humanoids), which are themselves still in draft. A realistic external-availability window is 2027-2028 for industrial partners and later for any consumer offering. Treat shorter timelines as marketing.

Further reading

Internal references on adjacent topics:

• NVIDIA Jetson Thor humanoid robot reference architecture for 2026 platforms — the dual-die alternative to Tesla’s single-SoC choice.
• Vision-Language-Action foundation models (pi-0, RT-2) for humanoids in 2026 — the model family behind Optimus’s policy approach.
• MoveIt 2 vs Tesseract for robot motion planning in 2026 — comparison of the open-source planners that Tesla’s custom MPC layer would replace.

External authoritative sources:

• Tesla AI Day 2022 humanoid presentation — original Optimus actuator and battery disclosures.
• IEEE Spectrum humanoid robotics coverage — independent reporting on humanoid actuator design trends.
• The Information’s robotics and Optimus reporting — teardown and development timeline reporting.
• ISO TC 299 robotics standards work — the standards body whose draft work governs humanoid commercial deployment.